Georgi Gerganov Reflects at llama.cpp 100K Stars: Local Model Failures Are Harness Bugs, Not Model Bugs

Simon Willison / Georgi Gerganov·medium signal

Simon Willison highlighted Gerganov's 100K-star reflection where the llama.cpp creator argues the main issues users face with local models 'mostly revolve around the harness and some intricacies around model chat templates and prompt construction' — not model quality. The pipeline from input to output involves components 'developed by different parties' that are 'not only fragile' but lack cohesion, with implementations 'with very high probability still broken in some subtle way along that chain.' This directly validates the Meta-Harness paper's thesis that harness engineering is now the bottleneck, not model capability.